We study adaptive detector selection for a heavily loaded 16-antenna, 8-user one-bit QPSK MIMO uplink with imperfect CSI under Rayleigh and correlated UMi channels, including an ill-conditioned LOS scenario. A contextual-bandit DQN selects among MMSE, ZF, MRC, Turbo-2, and Turbo-3 detectors from a 10-dimensional receiver-observable state, rewarded by an LDPC syndrome computed from the decoded codeword alone. All policies see identical channel, noise, and codeword realizations, and are scored by the observable syndrome clearance rate (SCR) and the unobservable true packet success (TPS). Linear MMSE is the TPS-best fixed detector in all six evaluated cells, so the selector delivers no TPS gain: it reaches that operating point from receiver-observable data alone, within 0.4 TPS points and without any channel-regime label. The fixed detectors span up to 44 TPS points, and the selector's cost adapts. The LOS cell exposes a two-layer observability limit: instantaneous channel descriptors separate the LOS regime only weakly, and the syndrome proxy itself saturates. Fixed Turbo-3 lifts SCR from 57% to 77% while TPS stays below 1% for every detector, a bias the syndrome-rewarded policy inherits. An embedded CRC-16 check tracked delivery over 81,000 packets with no undetected error, at a cost of 16 of the 24 information bits. Decoder-consistency rewards therefore track delivery only while the code's undetected-error rate is negligible.

deep reinforcement learning,detector selection,large-scale MIMO,LDPC syndrome,one-bit ADC,turbo detection